This repository is the implementation code of the paper "Sequential Dexterity: Chaining Dexterous Policies for Long-Horizon Manipulation"(Paper, Website, Presentation) by Yuanpei Chen*, Chen Wang*, Li Fei-Fei and C. Karen Liu at The Movement Lab and Stanford Vision and Learning Lab. In this repo, we provide our full implementation code of quick demo, training, evaluation, and real-world system.
- python 3.8
conda create -n seqdex python=3.8
conda activate seqdex
- IsaacGym (tested with
Preview Release 3/4andPreview Release 4/4). Follow the instruction to download the package.
tar -xvf IsaacGym_Preview_4_Package.tar.gz
cd isaacgym/python
pip install -e .
cd examples
(test installation) python examples/joint_monkey.py
- SeqDex
git clone https://github.com/sequential-dexterity/SeqDex.git
cd SeqDex
pip install -r requirements.txt
pip install -e .
- Trained checkpoint. Download from Link.
unzip seqdex_checkpoint.zip
mv checkpoint SeqDex/dexteroushandenvs/checkpoint
Here we provide a quick demo where you can interactively assign a LEGO building manual you prefer to the robot with the keyboard. The robot will load the trained checkpoints to perform the task. To get the best performance for this demo, the policy takes the full state information (object acceleration, motor velocity, ...) as inputs and is allowed to control the end-effector orientation, which is the policy before distillation for the real-world deployment. To start the demo, simply
cd SeqDex/dexteroushandenvs
python train_rlgames.py --task BlockAssemblyGUI --num_envs=1 --play
The control instructions are as follows:
Taking the Block Assembly as an example, each sub-task and their order are BlockAssemblySearch->BlockAssemblyOrient->BlockAssemblyGraspSim->BlockAssemblyInsertSim.
If you want to use the bi-directional optimization in our paper to train the BlockAssembly or ToolPositioning task, simply:
python scripts/bi-optimization.py --task=[BlockAssembly / ToolPositioning]
Since each sub-task takes about 1~2 days to train, it may take a long time for the whole process, so we also provide a way to train each sub-task individually. When training each sub-task, run this line in dexteroushandenvs folder:
python train_rlgames.py --task=[BlockAssemblySearch / BlockAssemblyOrient / BlockAssemblyGraspSim / BlockAssemblyInsertSim] --num_envs=1024
The trained model will be saved to runs folder, and the terminal state of the task will be saved to immediate_state. These terminal states are also used to train the transition feasibility function (see our paper), using the following command:
python policy_sequencing/tvalue_trainer.py --task=[BlockAssemblySearch / BlockAssemblyOrient / BlockAssemblyGraspSim / BlockAssemblyInsertSim]
To load a trained model and only perform inference (no training) in each sub-task, pass --play as an argument, and pass --checkpoint to specify the trained models which you want to load. Here is an example in BlockAssemblyGraspSim task:
python train_rlgames.py --task=BlockAssemblyGraspSim --checkpoint=./checkpoint/block_assembly/last_AllegroHandLegoTestPAISim_ep_19000_rew_1530.9819.pth --play --num_envs=256During evaluation in our paper, infer each sub-policy in the same order and calculate the overall success rate at the last task. Simply
python scripts/evaluation.py
It would save the terminal state of each sub-task in inference and finally count the success rate in the Insertion task.
You can also use the same program as the Quick demo for testing: assemble a LEGO. This requires adding the parameters --record_video and --lego_type to select the type of lego:
cd SeqDex/dexteroushand
python train_rlgames.py --task BlockAssemblyGUI --num_envs=1 --play --record_video --record_lego_type=0
We also save this results in /output_video/.
We thank the list of contributors from the Bi-DexHands.
Please cite Sequential Dexterity if you use this repository in your publications:
@article{chen2023sequential,
title={Sequential Dexterity: Chaining Dexterous Policies for Long-Horizon Manipulation},
author={Chen, Yuanpei and Wang, Chen and Fei-Fei, Li and Liu, C Karen},
journal={arXiv preprint arXiv:2309.00987},
year={2023}
}
Licensed under the MIT License